Susan Deane

I am a registered nurse and nurse educator with more than 20 years of bedside experience. Please see my attached comments, which address Questions 1, 3, 7, 9, 14, 21, and 26.

In summary, nurses will use generative and agentic AI-enabled devices, review their outputs, and staff many of the human checkpoints the discussion paper describes, yet the paper does not mention nurses. I ask CDRH to draw on nursing competency-validation models in the competency-based approach; to include nurses in benchmarking, clinical confirmation, and comparator selection when nurses do the work or use the device; to keep the care team connected to what patient-facing devices tell patients; to include nurses and nursing organizations in postmarket monitoring; and to design and evaluate human checkpoints for the nurses who will staff them.

These comments reflect my own views and do not represent any organization I am affiliated with.

Susan Deane, EdD, MSN, CNE, RN

FDA Comment — Susan Deane

FDA Comment — Docket FDA-2026-N-7874
Docket No. FDA-2026-N-7874 Considerations for the Regulation of Generative AI-Enabled
Medical Devices: Discussion Paper and Request for Feedback

Submitted by: Susan Deane, EdD, MSN, CNE, RN

I am submitting these comments as an individual. They reflect my own views and do not
represent any organization I am affiliated with.

Introduction
Thank you for the opportunity to comment on this discussion paper.

I am a registered nurse and nurse educator. I spent more than 20 years in bedside nursing,
including medical-surgical, intensive care, dialysis, and long-term care, and later served as
dean of a school of nursing and as interim provost at a university. My work now focuses on
preparing nurses to evaluate and use AI safely in clinical practice.

Nurses will be among the most frequent users of generative and agentic AI-enabled
devices. We will read their outputs, explain them to patients, catch their errors, and often
serve as the human checkpoint the paper describes. Yet the paper does not mention
nurses. My comments address where that gap matters most: Questions 7 and 9, 14, 3, 21,
and 26.

Questions 7 and 9: The competency-based approach
I support the competency-based approach of device benchmarking followed by clinical
confirmation. Evaluating an open-ended tool the way we evaluate clinicians is a sound
starting point.

However, the paper draws its model only from physician training: medical licensing, board
examinations, residency, and fellowship. Nursing has its own long-established system for
validating clinical competence, and it is directly relevant here:

Licensure through the NCLEX, which tests safe clinical judgment rather than
knowledge alone.
Orientation and preceptorship, where a new nurse performs specific tasks under
supervision, with independence granted step by step.

Page 1 of 5
FDA Comment — Susan Deane

Ongoing competency validation by employers, tied to specific skills, equipment, and
settings, and repeated over time.
Specialty certification for practice in areas such as critical care or nephrology.

This model fits devices well, because it validates safe performance of defined tasks in a
defined setting. That is closer to how a device is authorized for an intended use than
general medical training is.

Recommendations:

1. Draw on nursing competency-validation models, not physician training alone, as the
approach develops.
2. When nurses are intended or foreseeable users of a device, include nurses in
benchmarking and clinical confirmation, especially for element E.4 (communication
quality and user comprehension).
3. Test for automation bias with real users under realistic conditions, including the
workload and time pressure of an actual shift. A device that performs well in a quiet
test may be accepted without scrutiny on a busy unit.

Question 14: Who the device is compared against
The comparator should be the clinician who performs the task today. Question 14 asks
whether generalist or specialist physicians should set the performance standard. For
many of the tasks generative tools are being built for, neither is the right answer.

Nurses already carry out much of this work: triage advice by phone and patient portal,
care escalation decisions, patient education, symptom monitoring between visits, and
documentation. When a device takes on one of these tasks, its performance should be
measured against nurses doing that task, at the applicable standard of nursing care.

Question 14 also asks when human-AI team performance should be the basis for
evaluation. In most real deployments, a nurse will read, act on, or relay the device's output.
The team as deployed is what patients experience, so it should be evaluated directly.

Recommendations:

1. Select comparators by the task and the clinician who performs it today, including
nurses where nurses do the work.
2. Where a nurse will review or act on the output, evaluate human-AI team performance
with nurses in that role.

Page 2 of 5
FDA Comment — Susan Deane

3. In that evaluation, check whether the reviewing nurse has enough information and
time to recognize an incorrect output. A review step that cannot realistically catch
errors adds risk rather than safety.

Question 3: Patient-facing functions
The line between patient-facing and clinician-facing functions is less clean in practice
than the paper suggests. Nurses usually sit between the two. We answer the portal
message after the patient has already read the tool's advice, and we correct the
misunderstanding at the next visit or on the next call.

From my years in dialysis, I know how quickly patient-facing advice can matter. Advice
about fluid or potassium intake can carry serious consequences, and patients act on it
between treatments, before a nurse sees them.

I agree that patients deserve access to clinical information, and the paper is right not to
underestimate them. The safeguards should keep the care team connected to what the
patient was told.

Recommendations:

1. Require a clear escalation route from patient-facing devices to a reachable member of
the care team.
2. Make a record of what the device told the patient available to the clinicians who follow
up, so they are not working blind.
3. Test patient comprehension across health literacy levels, as element E.4 describes,
before authorization.

Question 21: Roles in postmarket monitoring
Nurses are well placed to detect when a deployed tool starts to go wrong. We are with
patients around the clock, and we see a tool's outputs across many patients and shifts. A
change in how a tool behaves is often visible to nurses before it shows up in any formal
analysis.

Those observations only help if they can be captured. Unless reporting is simple and built
into the workflow, what a nurse notices stays on the unit.

I agree with the paper that monitoring should not diffuse manufacturer accountability.
Nurses and other clinicians should report what they see; manufacturers should remain
responsible for acting on it.

Page 3 of 5
FDA Comment — Susan Deane

Recommendations:

1. Expect manufacturers to provide a simple, in-workflow way for frontline users,
including nurses, to flag a suspected incorrect or unsafe output.
2. Include nurses on the clinician review panels described in Section VI.A when nurses
use the device.
3. Invite professional nursing organizations to help develop reporting standards and
education.
4. Recognize education as part of the ecosystem. Nurses need preparation to recognize
AI errors, and nursing education is only beginning to provide it.

Questions 26 and 1: Agentic systems and human checkpoints
In many clinical workflows, the human checkpoint before an agentic system acts will be a
nurse. Element A.1 rightly asks whether a system complies with human-oversight
checkpoints before irreversible or high-consequence actions. But a checkpoint is only as
strong as the review behind it.

If a nurse is asked to approve dozens of AI-initiated actions in a shift, without seeing how
the system reached them, approval becomes a click. In my teaching, I train nurses to
pause and verify an AI recommendation before acting on it. That pause only works when
the system leaves room for it.

On Question 1, I support adding reversibility, the availability of downstream safeguards,
and time pressure to the risk framework. One caution: the downstream safeguard is often
a nurse. A nurse should count as a safeguard only when that nurse has the time,
information, and authority to intervene.

Recommendations:

1. Evaluate human checkpoints as they will actually be staffed, including the reviewer's
realistic workload.
2. At each checkpoint, show the reviewer the steps and sources behind the proposed
action, not only the action.
3. Make stopping or modifying an action as easy as approving it.
4. Treat a nurse as a safeguard in the risk framework only when the nurse can realistically
intervene.

Page 4 of 5
FDA Comment — Susan Deane

Summary
Nurses will use these devices, review their outputs, and stand at many of the checkpoints
this paper describes. I ask CDRH to:

Draw on nursing competency-validation models alongside physician training in the
competency-based approach.
Include nurses in benchmarking, clinical confirmation, and comparator selection when
nurses do the work or use the device.
Keep the care team connected to what patient-facing devices tell patients.
Give nurses a simple way to report suspected errors, and include nurses and nursing
organizations in postmarket monitoring.
Design and evaluate human checkpoints for the nurses who will staff them.

AI informs. The nurse decides. Regulation that reflects nursing practice will help make
that true at the bedside.

Thank you for your consideration.

Susan Deane, EdD, MSN, CNE, RN Nurse educator; founder, The Digital Stethoscope
susandeane119@gmail.com

Page 5 of 5